Fig. 1

The four types of learning considered in our model. Individuals have the task of selecting the highest-quality food item among a set of available items. Food quality cannot be perceived directly, but only on the basis of cues like colour or smell. The ‘true’ relationship between cues and quality is described by the ‘environmental profile’, a function \(Q\,(C)\) that can change between generations. Each individual harbours a neural network that predicts the relationship between cues and quality. This network is inherited but can change during the individual’s lifetime due to individual or social experience, as described below. The coloured curves illustrate the profiles (= the predictions of quality from cues) of two individuals, a ‘learner’ (\(Q_{\,L} (C)\), red curve) and a ‘demonstrator’ (\(Q_{\,D} (C)\), blue curve). A learning event is modelled as follows: first, a number of random cues are offered to the learner (the six vertical lines in the second row of panels, though 10 are utilised in our simulations). Second, the learner selects a cue in one of three ways: at random (\(C_{R}\), left panel), the cue the learner predicts to be associated with the highest quality (\(\hat{C}_{L}\), middle panel), or the cue the demonstrator predicts to be associated with the highest quality (\(\hat{C}_{D}\), right panel). In all three cases, the learner has a prediction P of the quality associated with the selected cue (\(P = Q_{L} (C_{R} ),\;Q_{L} (\hat{C}_{L} ),\;{\text{or }}Q_{L} (\hat{C}_{D} )\)). Third, the learner receives feedback on the correctness of the prediction by comparing it to the ‘target value’ T. In the first three types of learning, the learner is informed about the ‘true’ quality of the selected item (\(T = Q(C_{R} ),\;Q(\hat{C}_{L} ),\;{\text{or }}Q(\hat{C}_{D} )\)); in the fourth variant, the learner is informed about the quality predicted by the demonstrator (\(T = Q_{D} (\hat{C}_{D} )\)). Fourth, the learner’s neural network is updated by applying the Delta rule (see Methods) that incorporates the discrepancy between predicted value P and target value T. This update leads to a changed prediction profile of the learner (\(Q_{L}^{\prime } (C)\)), who uses this network as the starting point of the next learning episode or, after learning is complete, uses it to make foraging decisions.